Complement method of missing streamflow time-series data by means of Long and Short-Term Memory network

Streamflow data are important for river maintenance and water resources management. However, flow data may be partially missing due to various factors. These days, many studies have shown high applicability in the field of hydrology such as streamflow prediction of deep learning methods. Therefore, there is a possibility that the missing flow data can be complemented with high accuracy by means of the deep learning method. This study proposed a novel method that utilized deep learning to complement missing time-series flow data. Among the deep learning methods, this study selected the Long and Short-Term Memory (LSTM) network which are kinds of recurrent neural networks (RNN). LSTM may be able to learn long-term dependencies and has several types. This study implemented both Unidirectional LSTM (Uni-LSTM) and Bidirectional LSTM (Bi-LSTM). Uni-LSTM learns time series data in the forward direction. On the other hand, Bi-LSTM learns in the backward direction as well as in the forward direction. As a study area, the Kikuchi-River Basin, which is located in the Kyushu region, Japan was selected. To complement the hourly missing flow data, the hourly flow data and basin average precipitation data are used as the input. Here, in order to train the model, and evaluate the estimation accuracy, the missing data were artificially generated. The results showed the high applicability both of Uni-LSTM and Bi-LSTM to complement missing flow data.